This page documents Broccoli, the Bro client communications library. It allows you to create client sensors for the Bro intrusion detection system. Broccoli can speak a good subset of the Bro communication protocol, in particular, it can receive Bro IDs, send and receive Bro events, and send and receive event requests to/from peering Bros.
Contents
broccoli-config
Broccoli is the BRO Client COmmunications LIbrary. It allows you to write applications that speak the communication protocol of the Bro intrusion detection system.
Broccoli is free software under terms of the BSD license as given in the COPYING file distributed with its source code.
In this document, we assume that you are familiar with the basic concepts of Bro, so please first review the documentation/publications available from the Bro website if necessary.
Feedback, patches and bug reports are all welcome, please see http://www.bro.org/community for instructions on how to participate in the Bro community.
Having a single IDS on your network is good, but things become a lot more interesting when you can communicate information among multiple vantage points in your network. Bro agents can communicate with other Bro agents, sending and receiving events and other state information. In the Bro context this is particularly interesting because it means that you can build sophisticated policy-controlled distributed event management systems.
Broccoli enters the picture when it comes to integrating components that are not Bro agents themselves. Broccoli lets you create applications that can speak the Bro communication protocol. You can compose, send, request, and receive events. You can register your own event handlers. You can talk to other Broccoli applications or Bro agents – Bro agents cannot tell whether they are talking to another Bro or a Broccoli application. Broccoli allows you to integrate applications of your choosing into a distributed policy-controlled event management system. Broccoli is intended to be portable: it should build on Linux, the BSDs, Solaris, and Windows (in the MinGW environment).
Unlike other distributed IDSs, Bro does not assume a strict sensor-manager hierarchy in the information flow. Instead, Bro agents can request delivery of arbitrary events from other instances. When an event is triggered in a Bro agent, it checks whether any connected agents have requested notification of this event, and sends a copy of the event, including the event arguments. Recall that in Bro, an event handler is essentially a function defined in the Bro language, and an event materializes through invocation of an event handler. Each remote agent can define its own event handlers.
Broccoli applications will typically do one or more of the following:
The installation process will hopefully be painless: Broccoli is
installed from source using the usual ./configure <options> && make &&
make install
routine after extraction of the tarball.
Some relevant configuration options to pass to configure are:
--prefix=<DIR>
: sets the installation root to DIR.
The default is to install below /usr/local
.--enable-debug
: enables debugging output.
Please refer to the Configuring Debugging Output section for
details on configuring and using debugging output.--with-configfile=<FILE>
: use FILE as location of configuration
file. See the section on Configuration Files for more on this.--with-openssl=<DIR>
: use the OpenSSL installation below DIR.After installation, you’ll find the library in shared and static
versions in <prefix>/lib
, the header file for compilation in
<prefix>/include
.
broccoli-config
Similarly to many other software packages, the Broccoli distribution provides a script that you can use to obtain details about your Broccoli setup. The script currently provides the following flags:
--build
prints the name of the machine the build was
made on, when, and whether debugging support was enabled or not.--prefix
prints the directory in the filesystem
below which Broccoli was installed.--version
prints the version of the distribution
you have installed.--libs
prints the flags to pass to the
linker in order to link in the Broccoli library.--cflags
prints the flags to pass to the
compiler in order to properly include Broccoli’s header file.--config
prints the location of the system-wide
config file your installation will use.The --cflags
and --libs
flags are the suggested way of obtaining
the necessary information for integrating Broccoli into your build
environment. It is generally recommended to use broccoli-config
for
this purpose, rather than, say, develop new autoconf tests. If you
use the autoconf/automake tools, we recommend something along the
following lines for your configure
script:
dnl ##################################################
dnl # Check for Broccoli
dnl ##################################################
AC_ARG_WITH(broccoli-config,
AC_HELP_STRING(\[--with-broccoli-config=FILE], \[Use given broccoli-config]),
[ brocfg="$withval" ],
[ AC_PATH_GENERIC(broccoli,,
brocfg="broccoli-config",
AC_MSG_ERROR(Cannot find Broccoli: Is broccoli-config in path? Use more fertilizer?)) ])
broccoli_libs=`$brocfg --libs`
broccoli_cflags=`$brocfg --cflags`
AC_SUBST(broccoli_libs)
AC_SUBST(broccoli_cflags)``
You can then use the compiler/linker flags in your Makefile.in/ams by substituting in the values accordingly, which might look as follows:
CFLAGS = -W -Wall -g -DFOOBAR @broccoli_cflags@
LDFLAGS = -L/usr/lib/foobar @broccoli_libs@
Often you will want to make existing applications Bro-aware, that is,
instrument them so that they can send and receive Bro events at
appropriate moments in the execution flow. This will involve modifying
an existing code tree, so care needs to be taken to avoid unwanted side
effects. By protecting the instrumented code with #ifdef
/#endif
statements you can still build the original application, using the
instrumented source tree. The broccoli-config
script helps you in
doing so because it already adds -DBROCCOLI
to the compiler flags
reported when run with the --cflags
option:
> broccoli-config --cflags
-I/usr/local/include -I/usr/local/include -DBROCCOLI
So simply surround all inserted code with a preprocessor check for
BROCCOLI
and you will be able to build the original application as
soon as BROCCOLI
is not defined.
Time for some code. In the code snippets below we will introduce variables
whenever context requires them and not necessarily when C requires them.
In order to make the API known, include broccoli.h
:
#ifdef BROCCOLI #include <broccoli.h> #endif
Note
Broccoli’s memory management philosophy:
Broccoli generally does not release objects you allocate. The approach taken is “you clean up what you allocate.”
Broccoli requires global initialization before most of its other functions can be used. Generally, the way to initialize Broccoli is as follows:
bro_init(NULL);
The argument to bro_init()
provides optional initialization context,
and may be kept NULL
for normal use. If required, you may allocate a
BroCtx
structure locally, initialize it using bro_ctx_init()
,
fill in additional values as required and subsequently pass it to
bro_init()
:
BroCtx ctx; bro_ctx_init(&ctx); /* Make adjustments to the context structure as required...*/ bro_init(&ctx);
Note
The BroCtx
structure currently contains a set of five
different callback function pointers. These are required for
thread-safe operation of OpenSSL (Broccoli itself is thread-safe).
If you intend to use Broccoli in a multithreaded environment, you
need to implement functions and register them via the BroCtx
structure. The O’Reilly book “Network Security with OpenSSL” by
Viega et al. shows how to implement these callbacks.
Warning
You must call bro_init()
at the start of your
application. Undefined behavior may result if you don’t.
Broccoli declares a number of data types in broccoli.h
that you
should know about. The more complex ones are kept opaque, while you do
get access to the fields in the simpler ones. The full list is as
follows:
Simple signed and unsigned types: int, uint, uint16, uint32, uint64 and uchar.
Connection handles: BroConn, kept opaque.
Bro events: BroEvent, kept opaque.
Buffer objects: BroBuf, kept opaque. See also Using Dynamic Buffers.
Ports: BroPort for network ports, defined as follows:
typedef struct bro_port { uint64 port_num; /* port number in host byte order */ int port_proto; /* IPPROTO_xxx */ } BroPort;
Records: BroRecord, kept opaque. See also Handling Records.
Strings (character and binary): BroString, defined as follows:
typedef struct bro_string { uint32 str_len; uchar *str_val; } BroString;
BroStrings are mostly kept transparent for convenience; please have a look at the Broccoli API Reference.
Tables: BroTable, kept opaque. See also Handling Tables.
Sets: BroSet, kept opaque. See also Handling Sets.
IP Address: BroAddr, defined as follows:
typedef struct bro_addr { uint32 addr[4]; /* IP address in network byte order */ int size; /* Number of 4-byte words occupied in addr */ } BroAddr;
Both IPv4 and IPv6 addresses are supported, with the former occupying
only the first 4 bytes of the addr
array.
Subnets: BroSubnet, defined as follows:
typedef struct bro_subnet { BroAddr sn_net; /* IP address in network byte order */ uint32 sn_width; /* Length of prefix to consider. */ } BroSubnet;
You can use Broccoli to establish a connection to a remote Bro, or to create a Broccoli-enabled server application that other Bros will connect to (this means that in principle, you can also use Broccoli purely as middleware and have multiple Broccoli applications communicate directly).
In order to establish a connection to a remote Bro, you first obtain a
connection handle. You then use this connection handle to request
events, connect to the remote Bro, send events, etc. Connection handles
are pointers to BroConn
structures, which are kept opaque. Use
bro_conn_new()
or bro_conn_new_str()
to obtain a handle,
depending on what parameters are more convenient for you: the former
accepts the IP address and port number as separate numerical arguments,
the latter uses a single string to encode both, in “hostname:port”
format.
To write a Broccoli-enabled server, you first need to implement the
usual socket()
/ bind()
/ listen()
/ accept()
routine.
Once you have obtained a file descriptor for the new connection from
accept()
, you pass it to the third function that returns a
BroConn
handle, bro_conn_new_socket()
. The rest of the
connection handling then proceeds as in the client scenario.
All three calls accept additional flags for fine-tuning connection behaviour. These flags are:
BRO_CFLAG_NONE
: no functionality. Use when no flags are desired.BRO_CFLAG_RECONNECT
:
When using this option, Broccoli will attempt to reconnect to the peer
transparently after losing connectivity. Essentially whenever you try to
read from or write to the peer and its connection has broke down, a full
reconnect including complete handshaking is attempted. You can check
whether the connection to a peer is alive at any time using
bro_conn_alive()
.BRO_CFLAG_ALWAYS_QUEUE
:
When using this option, Broccoli will queue any events you send for
later transmission when a connection is currently down. Without using
this flag, any events you attempt to send while a connection is down
get dropped on the floor. Note that Broccoli maintains a maximum queue
size per connection so if you attempt to send lots of events while the
connection is down, the oldest events may start to get dropped
nonetheless. Again, you can check whether the connection is currently
okay by using bro_conn_alive()
.BRO_CFLAG_DONTCACHE
:
When using this option, Broccoli will ask the peer not to use caching
on the objects it sends to us. This is the default, and the flag need
not normally be used. It is kept to maintain backward compatibility.BRO_CFLAG_CACHE
:
When using this option, Broccoli will ask the peer to use caching on
the objects it sends to us. Caching is normally disabled.BRO_CFLAG_YIELD
:
When using this option, bro_conn_process_input()
processes at most
one event at a time and then returns.By obtaining a connection handle, you do not also establish a connection
right away. This is done using bro_conn_connect()
. The main reason
for this is to allow you to subscribe to events (using
bro_event_registry_add()
, see Receiving Events) before
establishing the connection. Upon returning from bro_conn_connect()
you are guaranteed to receive all instances of the event types you have
requested, while later on during the connection some time may elapse
between the issuing of a request for events and the processing of that
request at the remote end. Connections are established via TCP,
optionally using SSL encryption. See “Configuring Encrypted
Communication”, for more information on setting up encryption. The
port numbers Bro agents and Broccoli applications listen on can vary
from peer to peer.
Finally, bro_conn_delete()
terminates a connection and releases all
resources associated with it. You can create as many connections as you
like, to one or more peers. You can obtain the file descriptor of a
connection using bro_conn_get_fd()
:
char host_str[] = "bro.yourorganization.com"; int port = 1234; struct hostent *host; BroConn *bc; if (! (host = gethostbyname(host_str)) || ! (host->h_addr_list[0])) { /* Error handling -- could not resolve host */ } /* In this example, we obtain a connection handle, then register event handlers, and finally connect to the remote Bro. */ /* First obtain a connection handle: */ if (! (bc = bro_conn_new((struct in_addr*) host->h_addr_list[0], htons(port), BRO_CFLAG_NONE))) { /* Error handling - could not get connection handle */ } /* Register event handlers: */ bro_event_registry_add(bc, "foo", bro_foo_handler, NULL); /* ... */ /* Now connect to the peer: */ if (! bro_conn_connect(bc)) { /* Error handling - could not connect to remote Bro. */ } /* Send and receive events ... */ /* Disconnect from Bro and clean up connection */ bro_conn_delete(bc);
Or simply use the string-based version:
char host_str[] = "bro.yourcompany.com:1234"; BroConn *bc; /* In this example we don't request any events from the peer, but we ask it not to use the serialization cache. */ /* Again, first obtain a connection handle: */ if (! (bc = bro_conn_new_str(host_str, BRO_CFLAG_DONTCACHE))) { /* Error handling - could not get connection handle */ } /* Now connect to the peer: */ if (! bro_conn_connect(bc)) { /* Error handling - could not connect to remote Bro. */ } /* ... */
When you want to establish connections from multiple Broccoli applications with different purposes, the peer needs a means to understand what kind of application each connection belongs to. The real meaning of “kind of application” here is “sets of event types to request”, because depending on the class of an application, the peer will likely want to receive different types of events.
Broccoli lets you set the class of a connection using
bro_conn_set_class()
. When using this feature, you need to call that
function before issuing a bro_conn_connect()
since the class of a
connection is determined at connection startup:
if (! (bc = bro_conn_new_str(host_str, BRO_CFLAG_DONTCACHE))) { /* Error handling - could not get connection handle */ } /* Set class of this connection: */ bro_conn_set_class(bc, "syslog"); if (! bro_conn_connect(bc)) { /* Error handling - could not connect to remote Bro. */ }
If your peer is a Bro node, you need to match the chosen connection
class in the remote Bro’s Communication::nodes
configuration. See
Configuring event reception in Bro scripts, for how to do
this. Finally, in order to obtain the class of a connection as
indicated by the remote side, use bro_conn_get_peer_class()
.
In order to send an event to the remote Bro agent, you first create an empty event structure with the name of the event, then add parameters to pass to the event handler at the remote agent, and then send off the event.
Let’s assume we want to request a report of all connections a remote Bro
currently keeps state for that match a given destination port and host
name and that have amassed more than a certain number of bytes. The
idea is to send an event to the remote Bro that contains the query,
identifiable through a request ID, and have the remote Bro answer us
with remote_conn
events containing the information we asked for. The
definition of our requesting event could look as follows in the Bro
policy:
event report_conns(req_id: int, dest_host: string, dest_port: port, min_size: count);
First, create a new event:
BroEvent *ev; if (! (ev = bro_event_new("report_conns"))) { /* Error handling - could not allocate new event. */ }
Now we need to add parameters to the event. The sequence and types must
match the event handler declaration – check the Bro policy to make sure
they match. The function to use for adding parameter values is
bro_event_add_val()
. All values are passed as pointer arguments
and are copied internally, so the object you’re pointing to stays
unmodified at all times. You clean up what you allocate. In order to
indicate the type of the value passed into the function, you need to
pass a numerical type identifier along as well. Table-1 lists the
value types that Broccoli supports along with the type identifier and
data structures to point to.
Type | Type tag | Data type pointed to |
---|---|---|
Boolean | BRO_TYPE_BOOL |
int |
Integer value | BRO_TYPE_INT |
uint64 |
Counter (nonnegative integers) | BRO_TYPE_COUNT |
uint64 |
Enums (enumerated values) | BRO_TYPE_ENUM |
uint64 (see also description of bro_event_add_val() ’s type_name argument) |
Floating-point number | BRO_TYPE_DOUBLE |
double |
Timestamp | BRO_TYPE_TIME |
double (see also bro_util_timeval_to_double() and bro_util_current_time() ) |
Time interval | BRO_TYPE_INTERVAL |
double |
Strings (text and binary) | BRO_TYPE_STRING |
BroString (see also family of bro_string_xxx() functions) |
Network ports | BRO_TYPE_PORT |
BroPort , with the port number in host byte order |
IPv4/IPv6 address | BRO_TYPE_IPADDR |
BroAddr , with the addr member in network byte order and size member indicating the address family and number of 4-byte words of addr that are occupied (1 for IPv4 and 4 for IPv6) |
IPv4/IPv6 subnet | BRO_TYPE_SUBNET |
BroSubnet , with the sn_net member in network byte order |
Record | BRO_TYPE_RECORD |
BroRecord (see also the family of bro_record_xxx() functions and their explanation below) |
Table | BRO_TYPE_TABLE |
BroTable (see also the family of bro_table_xxx() functions and their explanation below) |
Set | BRO_TYPE_SET |
BroSet (see also the family of bro_set_xxx() functions and their explanation below) |
Knowing these, we can now compose a request_connections
event:
BroString dest_host; BroPort dest_port; uint32 min_size; int req_id = 0; bro_event_add_val(ev, BRO_TYPE_INT, NULL, &req_id); req_id++; bro_string_set(&dest_host, "desthost.destdomain.com"); bro_event_add_val(ev, BRO_TYPE_STRING, NULL, &dest_host); bro_string_cleanup(&dest_host); dest_port.port_num = 80; dest_port.port_proto = IPPROTO_TCP; bro_event_add_val(ev, BRO_TYPE_PORT, NULL, &dest_port); min_size = 1000; bro_event_add_val(ev, BRO_TYPE_COUNT, NULL, &min_size);
The third argument to bro_event_add_val()
lets you specify a
specialization of the types listed in Table-1. This is generally not
necessary except for one situation: when using BRO_TYPE_ENUM
. You
currently cannot define a Bro-level enum type in Broccoli, and thus when
sending an enum value, you have to specify the type of the enum along
with the value. For example, in order to add an instance of enum
transport_proto
defined in Bro’s init-bare.bro
, you would use:
int transport_proto = 2; /* ... */ bro_event_add_val(ev, BRO_TYPE_ENUM, "transport_proto", &transport_proto);
to get the equivalent of “udp” on the remote side. The same system is
used to point out type names when calling bro_event_set_val()
,
bro_record_add_val()
, bro_record_set_nth_val()
, and
bro_record_set_named_val()
.
All that’s left to do now is to send off the event. For this, use
bro_event_send()
and pass it the connection handle and the event.
The function returns TRUE
when the event could be sent right away or
if it was queued for later delivery. FALSE
is returned on error. If
the event gets queued, this does not indicate an error – likely the
connection was just not ready to send the event at this point. Whenever
you call bro_event_send()
, Broccoli attempts to send as much of an
existing event queue as possible. Again, the event is copied internally
to make it easier for you to send the same event repeatedly. You clean
up what you allocate:
bro_event_send(bc, ev); bro_event_free(ev);
Two other functions may be useful to you: bro_event_queue_length()
tells you how many events are currently queued, and
bro_event_queue_flush()
attempts to flush the current event queue
and returns the number of events that do remain in the queue after the
flush.
Note
you do not normally need to call this function, queue flushing is attempted every time you send an event.
Receiving events is a little more work because you need to
Each of these steps is explained in the following sections.
When Broccoli receives an event, it tries to dispatch the event to callbacks registered for that event type. The place where callbacks get registered is called the callback registry. Any callbacks registered for the arriving event’s name are invoked with the parameters shipped with the event. There are two styles of argument passing to the event callbacks. Which one is better suited depends on your application.
Each event argument is passed via a pointer to the callback. This makes best sense when you know the type of the event and of its arguments, because it provides you immediate access to arguments as when using a normal C function.
In order to register a callback with expanded argument passing, use
bro_event_registry_add()
and pass it the connection handle, the name
of the event for which you register the callback, the callback itself
that matches the signature of the BroEventFunc
type, and any user
data (or NULL
) you want to see passed to the callback on each
invocation. The callback’s type is defined rather generically as
follows:
typedef void (*BroEventFunc) (BroConn *bc, void *user_data, ...);
It requires a connection handle as its first argument and a pointer to
user-provided callback data as the second argument. Broccoli will pass
the connection handle of the connection on which the event arrived
through to the callback. BroEventFunc
’s are variadic, because each
callback you provide is directly invoked with pointers to the parameters
of the event, in a format directly usable in C. All you need to know is
what type to point to in order to receive the parameters in the right
layout. Refer to Table-1 again for a summary of those types. Record
types are more involved and are addressed in more detail in Handling
Records.
Note
Note that all parameters are passed to the
callback as pointers, even elementary types such as int
that
would normally be passed directly. Also note that Broccoli manages
the lifecycle of event parameters and therefore you do not have
to clean them up inside the event handler.
Continuing our example, we will want to process the connection reports
that contain the responses to our report_conns
event. Let’s assume
those look as follows:
event remote_conn(req_id: int, conn: connection);
The reply events contain the request ID so we can associate requests
with replies, and a connection record (defined in init-bare.bro
in Bro).
(It’d be nicer to report all replies in a single event but we’ll
ignore that for now.) For this event, our callback would look like
this:
void remote_conn_cb(BroConn *bc, void *user_data, int *req_id, BroRecord *conn);
Once more, you clean up what you allocate, and since you never allocated
the space these arguments point to, you also don’t clean them up.
Finally, we register the callback using bro_event_registry_add()
:
bro_event_registry_add(bc, "remote_conn", remote_conn_cb, NULL);
In this case we have no additional data to be passed into the callback,
so we use NULL
for the last argument. If you have multiple events
you are interested in, register each one in this fashion.
This is designed for situations when you have to determine how to handle different types of events at runtime, for example when writing language bindings or when implementing generic event handlers for multiple event types. The callback is passed a connection handle and the user data as above but is only passed one additional pointer, a BroEvMeta structure. This structure contains all metadata about the event, including its name, timestamp (in UTC) of creation, number of arguments, the arguments’ types (via type tags as listed in Table-1), and the arguments themselves.
In order to register a callback with compact argument passing, use
bro_event_registry_add_compact()
and pass it similar arguments as
you’d use with bro_event_registry_add()
. The callback’s type is
defined as follows:
typedef void (*BroCompactEventFunc) (BroConn *bc, void *user_data, BroEvMeta *meta);
Note
As before, Broccoli manages the lifecycle of event parameters. You do not have to clean up the BroEvMeta structure or any of its contents.
Below is sample code for extracting the arguments from the BroEvMeta structure, using our running example. This is still written with the assumption that we know the types of the arguments, but note that this is not a requirement for this style of callback:
void remote_conn_cb(BroConn *bc, void *user_data, BroEvMeta *meta) { int *req_id; BroRecord *rec; /* For demonstration, print out the event's name: */ printf("Handling a %s event.\n", meta->ev_name); /* Sanity-check the number of arguments: */ if (meta->ev_numargs != 2) { /* error */ } /* Sanity-check the argument types: */ if (meta->ev_args[0].arg_type != BRO_TYPE_INT) { /* error */ } if (meta->ev_args[1].arg_type != BRO_TYPE_RECORD) { /* error */ } req_id = (int *) meta->ev_args[0].arg_data; rec = (BroRecord *) meta->ev_args[1].arg_data; /* ... */ }
Finally, register the callback using
bro_event_registry_add_compact()
:
bro_event_registry_add_compact(bc, "remote_conn", remote_conn_cb, NULL);
At this point, Broccoli knows what to do with the requested events upon
arrival. What’s left to do is to let the remote Bro know that you would
like to receive the events for which you registered. If you haven’t yet
called bro_conn_connect()
, then there is nothing to do, since that
function will request the registered events anyway. Once connected, you
can still request events. To do so, call
bro_event_registry_request()
:
bro_event_registry_request(bc);
This mechanism also implies that no unrequested events will be delivered to us (and if that happened for whatever reason, the event would simply be dropped on the floor).
Note
At the moment you cannot unrequest events, nor can you request events based on predicates on the values of the events’ arguments.
At this point the remote Bro will start sending you the requested events once they are triggered. What is left to do is to read the arriving events from the connection and trigger dispatching them to the registered callbacks.
If you are writing a new Bro-enabled application, this is easy, and you
can choose among two approaches: polling explicitly via Broccoli’s API,
or using select()
on the file handle associated with a BroConn. The
former case is particularly straightforward; all you need to do is call
bro_conn_process_input()
, which will go off and check if any events
have arrived and if so, dispatch them accordingly. This function does
not block – if no events have arrived, then the call will return
immediately. For more fine-grained control over your I/O handling, you
will probably want to use bro_conn_get_fd()
to obtain the file
descriptor of your connection and then incorporate that in your standard
FD_SET
/select()
code. Once you have determined that data in fact
are ready to be read from the obtained file descriptor, you can then try
another bro_conn_process_input()
this time knowing that it’ll find
something to dispatch.
As a side note, if you don’t process arriving events frequently enough, then TCP’s flow control will start to slow down the sender until eventually events will queue up and be dropped at the sending end.
Broccoli supports record structures, i.e., types that pack a set of
values together, placing each value into its own field. In Broccoli, the
way you handle records is somewhat similar to events: after creating an
empty record (of opaque type BroRecord
), you can iteratively add
fields and values to it. The main difference is that you must specify a
field name with the value; each value in a record can be identified both
by position (a numerical index starting from zero), and by field name.
You can retrieve vals in a record by field index or field name. You can
also reassign values. There is no explicit, IDL-style definition of
record types. You define the type of a record implicitly by the sequence
of field names and the sequence of the types of the values you put into
the record.
Note that all fields in a record must be assigned before it can be shipped.
The API for record composition consists of bro_record_new()
,
bro_record_free()
, bro_record_add_val()
,
bro_record_set_nth_val()
, and bro_record_set_named_val()
.
On records that use field names, the names of individual fields can be
extracted using bro_record_get_nth_name()
. Extracting values from a
record is done using bro_record_get_nth_val()
and
bro_record_get_named_val()
. The former allows numerical indexing of
the fields in the record, the latter provides name-based lookups. Both
need to be passed the record you want to extract a value from, the index
or name of the field, and either a pointer to an int holding a
BRO_TYPE_xxx value (see again Table-1 for a summary of those types) or
NULL
. The pointer, if not NULL
, serves two purposes: type
checking and type retrieval. Type checking is performed if the value of
the int upon calling the functions is not BRO_TYPE_UNKNOWN. The type tag
of the requested record field then has to match the type tag stored in
the int, otherwise NULL
is returned. If the int stores
BRO_TYPE_UNKNOWN upon calling, no type-checking is performed. In both
cases, the actual type of the requested record field is returned in
the int pointed to upon return from the function. Since you have no
guarantees of the type of the value upon return if you pass NULL
as
the int pointer, this is a bad idea and either BRO_TYPE_UNKNOWN or
another type value should always be used.
For example, you could extract the value of the record field “label”, which we assume should be a string, in the following ways:
BroRecord *rec = /* obtained somehow */ BroString *string; int type; /* --- Example 1 --- */ type = BRO_TYPE_STRING; /* Use type-checking, will not accept other types */ if (! (string = bro_record_get_named_val(rec, "label", &type))) { /* Error handling, either there's no field of that value or the value is not of BRO_TYPE_STRING. The actual type is now stored in "type". */ } /* --- Example 2 --- */ type = BRO_TYPE_UNKNOWN; /* No type checking, just report the existent type */ if (! (string = bro_record_get_named_val(rec, "label", &type))) { /* Error handling, no field of that name exists. */ } printf("The type of the value in field 'label' is %i\n", type); /* --- Example 3 --- */ if (! (string = bro_record_get_named_val(rec, "label", NULL))) { /* Error handling, no field of that name exists. */ } /* We now have a value, but we can't really be sure of its type */
Record fields can be records, for example in the case of Bro’s standard
connection record type. In this case, in order to get to a nested
record, you use BRO_TYPE_RECORD
:
void remote_conn_cb(BroConn *bc, int *req_id, BroRecord *conn) { BroRecord *conn_id; int type = BRO_TYPE_RECORD; if ( ! (conn_id = bro_record_get_named_val(conn, "id", &type))) { /* Error handling */ } }
Broccoli supports Bro-style tables, i.e., associative containers that map instances of a key type to an instance of a value type. A given key can only ever point to a single value. The key type can be composite, i.e., it may consist of an ordered sequence of different types, or it can be direct, i.e., consisting of a single type (such as an integer, a string, or a record).
The API for table manipulation consists of bro_table_new()
bro_table_free()
, bro_table_insert()
, bro_table_find()
,
bro_table_get_size()
, bro_table_get_types()
, and
bro_table_foreach()
.
Tables are handled similarly to records in that typing is determined
dynamically by the initial key/value pair inserted. The resulting types
can be obtained via bro_table_get_types()
. Should the types not
have been determined yet, BRO_TYPE_UNKNOWN
will result. Also, as
with records, values inserted into the table are copied internally, and
the ones passed to the insertion functions remain unaffected.
In contrast to records, table entries can be iterated. By passing a
function of signature BroTableCallback()
and a pointer to data of
your choosing, bro_table_foreach()
will invoke the given function
for each key/value pair stored in the table. Return TRUE
to keep
the iteration going, or FALSE
to stop it.
Note
The main thing to know about Broccoli’s tables is how to use
composite key types. To avoid additional API calls, you may treat
composite key types exactly as records, though you do not need to use
field names when assigning elements to individual fields. So in order
to insert a key/value pair, you create a record with the needed items
assigned to its slots, and use this record as the key object. In
order to differentiate composite index types from direct ones
consisting of a single record, use BRO_TYPE_LIST
as the type of
the record, as opposed to BRO_TYPE_RECORD
. Broccoli will then
know to interpret the record as an ordered sequence of items making
up a composite element, not a regular record.
brotable.c
in the test/
subdirectory of the Broccoli tree
contains an extensive example of using tables with composite as well as
direct indexing types.
Sets are essentially tables with void value types. The API for set
manipulation consists of bro_set_new()
, bro_set_free()
,
bro_set_insert()
, bro_set_find()
, bro_set_get_size()
,
bro_set_get_type()
, and bro_set_foreach()
.
You will often find that you would like to connect data with a
BroConn
. Broccoli provides an API that lets you associate data items
with a connection handle through a string-based key-value registry. The
functions of interest are bro_conn_data_set()
,
bro_conn_data_get()
, and bro_conn_data_del()
. You need to
provide a string identifier for a data item and can then use that string
to register, look up, and remove the associated data item. Note that
there is currently no mechanism to trigger a destructor function for
registered data items when the Bro connection is terminated. You
therefore need to make sure that all data items that you do not have
pointers to via some other means are properly released before calling
bro_disconnect()
.
Imagine you have instrumented the mother of all server applications. Building it takes forever, and every now and then you need to change some of the parameters that your Broccoli code uses, such as the host names of the Bro agents to talk to. To allow you to do this quickly, Broccoli comes with support for configuration files. All you need to do is change the settings in the file and restart the application (we’re considering adding support for volatile configuration items that are read from the file every time they are requested).
A configuration is read from a single configuration file. This file can be read from different locations. Broccoli searches in this order for the config file:
BROCCOLI_CONFIG_FILE
environment
variable.~/.broccoli.conf
.broccoli-config --config
.Note
BROCCOLI_CONFIG_FILE
or ~/.broccoli.conf
will only be
used if it is a regular file, not executable, and neither group nor
others have any permissions on the file. That is, the file’s
permissions must look like -rw-------
or -r--------
.
In the configuration file, a #
anywhere starts a comment that runs to
the end of the line. Configuration items are specified as key-value
pairs:
# This is the Broccoli system-wide configuration file.
#
# Entries are of the form <identifier> <value>, where the
# identifier is a sequence of letters, and value can be a string
# (including whitespace), and floating point or integer numbers.
# Comments start with a "#" and go to the end of the line. For
# boolean values, you may also use "yes", "on", "true", "no",
# "off", or "false". Strings may contain whitespace, but need
# to be surrounded by double quotes '"'.
#
# Examples:
#
Foo/PeerName mybro.securesite.com
Foo/PortNum 123
Bar/SomeFloat 1.23443543
Bar/SomeLongStr "Hello World"
You can also have multiple sections in your configuration. Your application can select a section as the current one, and queries for configuration settings will then only be answered with values specified in that section. A section is started by putting its name (no whitespace please) between square brackets. Configuration items positioned before the first section title are in the default domain and will be used by default:
# This section contains all settings for myapp.
[ myapp ]
You can name identifiers any way you like, but to keep things organized
it is recommended to keep a namespace hierarchy similar to the file
system. In the code, you can query configuration items using
bro_conf_get_str()
, bro_conf_get_int()
, and
bro_conf_get_dbl()
. You can switch between sections using
bro_conf_set_domain()
.
Broccoli provides an API for dynamically allocatable, growable,
shrinkable, and consumable buffers with BroBuf
. You may or may not
find this useful – Broccoli mainly provides this feature in
broccoli.h
because these buffers are used internally anyway and
because they are a typical case of something that people implement
themselves over and over again, for example to collect a set of data
before sending it through a file descriptor, etc.
The buffers work as follows. The structure implementing a buffer is
called BroBuf
, and is initialized to a default size when
created via bro_buf_new()
and released using bro_buf_free()
.
Each BroBuf
has a content pointer that points to an arbitrary
location between the start of the buffer and the first byte after the
last byte currently used in the buffer (see buf_off
in the
illustration below). The content pointer can seek to arbitrary
locations, and data can be copied from and into the buffer, adjusting
the content pointer accordingly. You can repeatedly append data to the end
of the buffer’s used contents using bro_buf_append()
.
<---------------- allocated buffer space ------------>
<======== used buffer space ========> ^
^ ^ ^ |
| | | |
buf buf_ptr buf_off buf_len
Have a look at the following functions for the details:
bro_buf_new()
, bro_buf_free()
, bro_buf_append()
,
bro_buf_consume()
, bro_buf_reset()
, bro_buf_get()
,
bro_buf_get_end()
, bro_buf_get_size()
,
bro_buf_get_used_size()
, bro_buf_ptr_get()
,
bro_buf_ptr_tell()
, bro_buf_ptr_seek()
, bro_buf_ptr_check()
,
and bro_buf_ptr_read()
.
Encrypted communication between Bro peers takes place over an SSL connection in which both endpoints of the connection are authenticated. This requires at least some PKI in the form of a certificate authority (CA) which you use to issue and sign certificates for your Bro peers. To facilitate the SSL setup, each peer requires three documents: a certificate signed by the CA and containing the public key, the corresponding private key, and a copy of the CA’s certificate.
The OpenSSL command line tool openssl
can be used to create all
files necessary, but its unstructured arguments and poor documentation
make it a pain to use and waste lots of people a lot of time [1].
For an alternative tool to create SSL certificates for secure Bro/Broccoli
communication, see the create-cert
tool available at
ftp://ee.lbl.gov/create-cert.tar.gz.
In order to enable encrypted communication for your Broccoli
application, you need to put the CA certificate and the peer certificate
in the /broccoli/ca_cert
and /broccoli/host_cert
keys,
respectively, in the configuration file. Optionally, you can store the
private key in a separate file specified by /broccoli/host_key
. To
quickly enable/disable a certificate configuration, the
/broccoli/use_ssl
key can be used.
Note
This is where you configure whether to use encrypted or unencrypted connections.
If the /broccoli/use_ssl
key is present and set to one of “yes”,
“true”, “on”, or 1, then SSL will be used and an incorrect or missing
certificate configuration will cause connection attempts to fail. If
the key’s value is one of “no”, “false”, “off”, or 0, then in no case
will SSL be used and connections will always be cleartext.
If the /broccoli/use_ssl
key is not present, then SSL will be
used if a certificate configuration is found, and invalid
certificates will cause the connection to fail. If no certificates
are configured, cleartext connections will be used.
In no case does an SSL-enabled setup ever fall back to a cleartext one.
/broccoli/use_ssl yes
/broccoli/ca_cert <path>/ca_cert.pem
/broccoli/host_cert <path>/bro_cert.pem
/broccoli/host_key <path>/bro_cert.key
In a Bro policy, you need to load the frameworks/communication/listen.bro
script and redef Communication::listen_ssl=T
,
ssl_ca_certificate
, and ssl_private_key
, defined in init-bare.bro
:
@load frameworks/communication/listen redef Communication::listen_ssl=T; redef ssl_ca_certificate = "<path>/ca_cert.pem"; redef ssl_private_key = "<path>/bro.pem";
By default, you will be prompted for the passphrase for the private key matching the public key in your agent’s certificate. Depending on your application’s user interface and deployment, this may be inappropriate. You can store the passphrase in the config file as well, using the following identifier:
/broccoli/host_pass foobar
Warning
Make sure that access to your configuration is restricted.
If you provide the passphrase this way, it is obviously essential to have restrictive permissions on the configuration file. Broccoli partially enforces this. Please refer to the section on Configuration Files for details.
Before a remote Bro will accept your connection and your events, it needs to have its policy configured accordingly:
Load frameworks/communication/listen
, and redef the boolean variable
Communication::listen_ssl
depending on whether you want to have
encrypted or cleartext communication. Obviously, encrypting the event
exchange is recommended and cleartext should only be used for early
experimental setups. See below for details on how to set up encrypted
communication via SSL.
You need to find a port to use for the Bros and Broccoli applications
that will listen for connections. Every such agent can use a
different port, though default ports are provided in the Bro
policies. To change the port the Bro agent will be listening on from
its default, redefine the Communication::listen_port
. Have a
look at these policies as well as
base/frameworks/communication/main.bro
for the default values.
Here is the policy for the unencrypted case:
@load frameworks/communication/listen redef Communication::listen_port = 12345/tcp;
Including the settings for the cryptographic files introduced in the previous section, here is the encrypted one:
@load frameworks/communication/listen redef Communication::listen_ssl = T; redef Communication::listen_port = 12345/tcp; redef ssl_ca_certificate = "<path>/ca_cert.pem"; redef ssl_private_key = "<path>/bro.pem";
The policy controlling which peers a Bro agent will communicate with
and how this communication will happen are defined in the
Communication::nodes
table defined in
base/frameworks/communication/main.bro
. This table contains
entries of type Node
, whose members mostly provide default values
so you do not need to define everything. You need to come up with a
tag for the connection under which it can be found in the table (a
creative one would be “broccoli”), the IP address of the peer, the
pattern of names of the events the Bro will accept from you, whether
you want Bro to connect to your machine on startup or not, if so, a
port to connect to (default is Communication::default_port
also defined in
base/frameworks/communication/main.bro
), a retry timeout,
whether to use SSL, and the class of a connection as set on the
Broccoli side via bro_conn_set_class()
.
An example could look as follows:
redef Communication::nodes += { ["broping"] = [$host = 127.0.0.1, $class="broping", $events = /ping/, $connect=F, $ssl=F] };
This example is taken from broping.bro
, the policy the remote Bro
must run when you want to use the broping
tool explained in the
section on test programs below. It will allow an agent on the
local host to connect and send “ping” events. Our Bro will not
attempt to connect, and incoming connections will be expected in
cleartext.
If your Broccoli installation was configured with --enable-debug
,
Broccoli will report two kinds of debugging information:
Both are enabled by default, but can be adjusted in two ways.
/broccoli/debug_messages
and /broccoli/debug_calltrace
to on
/off
to enable/disable
the corresponding output.bro_debug_calltrace
and bro_debug_messages
to 1/0 at any time
to enable/disable the corresponding output.By default, debugging output is inactive (even with debug support
compiled in). You need to enable it explicitly either in your code by
assigning 1 to bro_debug_calltrace
and bro_debug_messages
or by
enabling it in the configuration file.
The Broccoli distribution comes with a few small test programs, located
in the test/
directory of the tree. The most notable one is
broping
[2], a mini-version of ping. It sends “ping” events to a
remote Bro agent, expecting “pong” events in return. It operates in two
flavours: one uses atomic types for sending information across, and the
other one uses records. The Bro agent you want to ping needs to run
either the broping.bro
or broping-record.bro
policies. You can
find these in the test/
directory of the source tree, and in
<prefix>/share/broccoli
in the installed version. broping.bro
is
shown below. By default, pinging a Bro on the same machine is
configured. If you want your Bro to be pinged from another machine, you
need to update the Communication::nodes
variable accordingly:
@load frameworks/communication/listen; global ping_log = open_log_file("ping"); redef Communication::nodes += { ["broping"] = [$host = 127.0.0.1, $events = /ping/, $connect=F, $retry = 60 secs, $ssl=F] }; event ping(src_time: time, seq: count) { event pong(src_time, current_time(), seq); } event pong(src_time: time, dst_time: time, seq: count) { print ping_log, fmt("ping received, seq %d, %f at src, %f at dest, one-way: %f", seq, src_time, dst_time, dst_time-src_time); }
broping
sends ping events to Bro. Bro accepts those because they are
configured accordingly in the nodes table. As shown in the
policy, ping events trigger pong events, and broccoli
requests
delivery of all pong events back to it. When running broping
,
you’ll see something like this:
> ./test/broping pong event from 127.0.0.1: seq=1, time=0.004700/1.010303 s pong event from 127.0.0.1: seq=2, time=0.053777/1.010266 s pong event from 127.0.0.1: seq=3, time=0.006435/1.010284 s pong event from 127.0.0.1: seq=4, time=0.020278/1.010319 s pong event from 127.0.0.1: seq=5, time=0.004563/1.010187 s pong event from 127.0.0.1: seq=6, time=0.005685/1.010393 s
The API documentation describes Broccoli’s public C interface.